forked from gmalivenko/pytorch2keras
-
Notifications
You must be signed in to change notification settings - Fork 0
/
squeezenext.py
407 lines (345 loc) · 12.4 KB
/
squeezenext.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
"""
Model from https://github.com/osmr/imgclsmob/tree/master/pytorch/models
"""
import numpy as np
import torch
from torch.autograd import Variable
from pytorch2keras.converter import pytorch_to_keras
import torchvision
import os
import torch.nn as nn
import torch.nn.init as init
class SqnxtConv(nn.Module):
"""
SqueezeNext specific convolution block.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
kernel_size : int or tuple/list of 2 int
Convolution window size.
stride : int or tuple/list of 2 int
Strides of the convolution.
padding : int or tuple/list of 2 int, default (0, 0)
Padding value for convolution layer.
"""
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
padding=(0, 0)):
super(SqnxtConv, self).__init__()
self.conv = nn.Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding)
self.bn = nn.BatchNorm2d(num_features=out_channels)
self.activ = nn.ReLU(inplace=True)
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
x = self.activ(x)
return x
class SqnxtUnit(nn.Module):
"""
SqueezeNext unit.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
stride : int or tuple/list of 2 int
Strides of the convolution.
"""
def __init__(self,
in_channels,
out_channels,
stride):
super(SqnxtUnit, self).__init__()
if stride == 2:
reduction_den = 1
self.resize_identity = True
elif in_channels > out_channels:
reduction_den = 4
self.resize_identity = True
else:
reduction_den = 2
self.resize_identity = False
self.conv1 = SqnxtConv(
in_channels=in_channels,
out_channels=(in_channels // reduction_den),
kernel_size=1,
stride=stride)
self.conv2 = SqnxtConv(
in_channels=(in_channels // reduction_den),
out_channels=(in_channels // (2 * reduction_den)),
kernel_size=1,
stride=1)
self.conv3 = SqnxtConv(
in_channels=(in_channels // (2 * reduction_den)),
out_channels=(in_channels // reduction_den),
kernel_size=(1, 3),
stride=1,
padding=(0, 1))
self.conv4 = SqnxtConv(
in_channels=(in_channels // reduction_den),
out_channels=(in_channels // reduction_den),
kernel_size=(3, 1),
stride=1,
padding=(1, 0))
self.conv5 = SqnxtConv(
in_channels=(in_channels // reduction_den),
out_channels=out_channels,
kernel_size=1,
stride=1)
if self.resize_identity:
self.identity_conv = SqnxtConv(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
stride=stride)
self.activ = nn.ReLU(inplace=True)
def forward(self, x):
if self.resize_identity:
identity = self.identity_conv(x)
else:
identity = x
identity = self.activ(identity)
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
x = self.conv4(x)
x = self.conv5(x)
x = x + identity
x = self.activ(x)
return x
class SqnxtInitBlock(nn.Module):
"""
SqueezeNext specific initial block.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
"""
def __init__(self,
in_channels,
out_channels):
super(SqnxtInitBlock, self).__init__()
self.conv = SqnxtConv(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=7,
stride=2,
padding=1)
self.pool = nn.MaxPool2d(
kernel_size=3,
stride=2,
ceil_mode=True)
def forward(self, x):
x = self.conv(x)
x = self.pool(x)
return x
class SqueezeNext(nn.Module):
"""
SqueezeNext model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
channels : list of list of int
Number of output channels for each unit.
init_block_channels : int
Number of output channels for the initial unit.
final_block_channels : int
Number of output channels for the final block of the feature extractor.
in_channels : int, default 3
Number of input channels.
num_classes : int, default 1000
Number of classification classes.
"""
def __init__(self,
channels,
init_block_channels,
final_block_channels,
in_channels=3,
num_classes=1000):
super(SqueezeNext, self).__init__()
self.features = nn.Sequential()
self.features.add_module("init_block", SqnxtInitBlock(
in_channels=in_channels,
out_channels=init_block_channels))
in_channels = init_block_channels
for i, channels_per_stage in enumerate(channels):
stage = nn.Sequential()
for j, out_channels in enumerate(channels_per_stage):
stride = 2 if (j == 0) and (i != 0) else 1
stage.add_module("unit{}".format(j + 1), SqnxtUnit(
in_channels=in_channels,
out_channels=out_channels,
stride=stride))
in_channels = out_channels
self.features.add_module("stage{}".format(i + 1), stage)
self.features.add_module('final_block', SqnxtConv(
in_channels=in_channels,
out_channels=final_block_channels,
kernel_size=1,
stride=1))
in_channels = final_block_channels
self.features.add_module('final_pool', nn.AvgPool2d(
kernel_size=7,
stride=1))
self.output = nn.Linear(
in_features=in_channels,
out_features=num_classes)
self._init_params()
def _init_params(self):
for name, module in self.named_modules():
if isinstance(module, nn.Conv2d):
init.kaiming_uniform_(module.weight)
if module.bias is not None:
init.constant_(module.bias, 0)
def forward(self, x):
x = self.features(x)
x = x.view(x.size(0), -1)
x = self.output(x)
return x
def get_squeezenext(version,
width_scale,
model_name=None,
pretrained=False,
root=os.path.join('~', '.torch', 'models'),
**kwargs):
"""
Create SqueezeNext model with specific parameters.
Parameters:
----------
version : str
Version of SqueezeNet ('23' or '23v5').
width_scale : float
Scale factor for width of layers.
model_name : str or None, default None
Model name for loading pretrained model.
pretrained : bool, default False
Whether to load the pretrained weights for model.
ctx : Context, default CPU
The context in which to load the pretrained weights.
root : str, default '~/.torch/models'
Location for keeping the model parameters.
"""
init_block_channels = 64
final_block_channels = 128
channels_per_layers = [32, 64, 128, 256]
if version == '23':
layers = [6, 6, 8, 1]
elif version == '23v5':
layers = [2, 4, 14, 1]
else:
raise ValueError("Unsupported SqueezeNet version {}".format(version))
channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)]
if width_scale != 1:
channels = [[int(cij * width_scale) for cij in ci] for ci in channels]
init_block_channels = int(init_block_channels * width_scale)
final_block_channels = int(final_block_channels * width_scale)
net = SqueezeNext(
channels=channels,
init_block_channels=init_block_channels,
final_block_channels=final_block_channels,
**kwargs)
if pretrained:
if (model_name is None) or (not model_name):
raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.")
import torch
from .model_store import get_model_file
net.load_state_dict(torch.load(get_model_file(
model_name=model_name,
local_model_store_dir_path=root)))
return net
def sqnxt23_w1(**kwargs):
"""
1.0-SqNxt-23 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.torch/models'
Location for keeping the model parameters.
"""
return get_squeezenext(version="23", width_scale=1.0, model_name="sqnxt23_w1", **kwargs)
def sqnxt23_w3d2(**kwargs):
"""
0.75-SqNxt-23 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.torch/models'
Location for keeping the model parameters.
"""
return get_squeezenext(version="23", width_scale=1.5, model_name="sqnxt23_w3d2", **kwargs)
def sqnxt23_w2(**kwargs):
"""
0.5-SqNxt-23 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.torch/models'
Location for keeping the model parameters.
"""
return get_squeezenext(version="23", width_scale=2.0, model_name="sqnxt23_w2", **kwargs)
def sqnxt23v5_w1(**kwargs):
"""
1.0-SqNxt-23v5 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.torch/models'
Location for keeping the model parameters.
"""
return get_squeezenext(version="23v5", width_scale=1.0, model_name="sqnxt23v5_w1", **kwargs)
def sqnxt23v5_w3d2(**kwargs):
"""
0.75-SqNxt-23v5 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.torch/models'
Location for keeping the model parameters.
"""
return get_squeezenext(version="23v5", width_scale=1.5, model_name="sqnxt23v5_w3d2", **kwargs)
def sqnxt23v5_w2(**kwargs):
"""
0.5-SqNxt-23v5 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615.
Parameters:
----------
pretrained : bool, default False
Whether to load the pretrained weights for model.
root : str, default '~/.torch/models'
Location for keeping the model parameters.
"""
return get_squeezenext(version="23v5", width_scale=2.0, model_name="sqnxt23v5_w2", **kwargs)
if __name__ == '__main__':
max_error = 0
for i in range(10):
model = sqnxt23_w1()
for m in model.modules():
m.training = False
input_np = np.random.uniform(0, 1, (1, 3, 224, 224))
input_var = Variable(torch.FloatTensor(input_np))
output = model(input_var)
k_model = pytorch_to_keras(model, input_var, (3, 224, 224,), verbose=True)
pytorch_output = output.data.numpy()
keras_output = k_model.predict(input_np)
error = np.max(pytorch_output - keras_output)
print(error)
if max_error < error:
max_error = error
print('Max error: {0}'.format(max_error))